On Cones of Nonnegative Quartic Formszhangs/Reports/2012_JLZ.pdfthe quadratic models. Though...
Transcript of On Cones of Nonnegative Quartic Formszhangs/Reports/2012_JLZ.pdfthe quadratic models. Though...
On Cones of Nonnegative Quartic Forms
Bo JIANG ∗ Zhening LI † Shuzhong ZHANG ‡
September 2, 2011
Abstract
Historically, much of the theory and practice in nonlinear optimization has revolved around
the quadratic models. Though quadratic functions are nonlinear polynomials, they are well
structured and easy to deal with. Limitations of the quadratics, however, become increasingly
binding as higher degree nonlinearity is imperative in modern applications of optimization. In
the recent years, one observes a surge of research activities in polynomial optimization, and
modeling with quartic or higher order polynomial functions has been more commonly accepted.
On the theoretical side, there are also major recent progresses on polynomial functions and
optimization. For instance, Ahmadi et al. [2] proved that checking the convexity of a quartic
polynomial function is strongly NP-hard in general, which settles a long-standing open question.
In this paper we proceed to studying six fundamentally important convex cones of quartic
functions in the space of symmetric quartic tensors, including the cone of nonnegative quartic
polynomials, the sum of squared polynomials, the convex quartic polynomials, and the sum of
fourth powered polynomials. It turns out that these convex cones coagulate into a chain in
decreasing order. The complexity status of these cones is sorted out as well. Finally, potential
applications of the new results to solve highly nonlinear and/or combinatorial optimization
problems are discussed.
Keywords: cone of polynomial functions; nonnegative quartic forms; sum of squares; sos-
convexity.
Mathematics Subject Classification 2010: 15A69, 08A40, 90C25, 49N15.
∗Industrial and Systems Engineering Program, University of Minnesota, Minneapolis, MN 55455. Email:
[email protected]†Department of Mathematics, Shanghai University, Shanghai, 200444, China. Email: [email protected]‡Industrial and Systems Engineering Program, University of Minnesota, Minneapolis, MN 55455. Email:
[email protected]. (On leave from Department of Systems Engineering and Engineering Management, The Chinese
University of Hong Kong, Shatin, Hong Kong. Email: [email protected])
1
1 Introduction
Checking the convexity of a quadratic function boils down to testing the positive semidefiniteness
of its Hessian matrix in the domain. Since the Hessian matrix is constant, the test can be done
easily. A natural question thus arises:
Given a fourth degree polynomial function in n variables, can one still easily tell if the
function is convex or not?
This simple-looking question was first put forward by Shor [28] in 1992, which turned out later to be
a very challenging question to answer. For almost two decades, the question remained open. Only
until recently Ahmadi et al. [2] proved that checking the convexity of a general quartic polynomial
function is actually strongly NP-hard. Their result not only settled this particular open problem,
but also helped to highlight a crucial difference between quartic and quadratic polynomials, which
makes the study of quartic polynomials all the more compelling and interesting.
On the practical side, quartic polynomial optimization has a wide spectrum of applications, includ-
ing sensor network localization [7], MIMO radar waveform optimization [12], portfolio management
with high moments information [20], quantum entanglement problem [13], among many others. This
has stimulated a burst of recent research activities with regard to quartic polynomial optimization.
Due to the NP-hardness of quartic polynomial optimization models (see e.g. [25, 15, 23]), there is a
considerable amount of recent research work devoted to the approximation algorithms for solving
various quartic polynomial optimization models. Luo and Zhang [25] proposed an approximation
algorithm for optimization of a quartic polynomial with quadratic constraints. Ling et al. [24] con-
sidered a special quartic optimization model, which is to maximize a bi-quadratic function over two
spheres. He et al. [15, 16] extended the study to arbitrary degree polynomials. So [35] improved
some of the approximation bounds presented in [15]. For a comprehensive survey on the topic,
one may refer to the recent Ph.D. thesis of Li [23]. Another well studied approach to cope with
general polynomial optimization problems is the so-called SOS method proposed by Lasserre [21]
and Parillo [29]. Theoretically, it can solve any general polynomial optimization model to any given
accuracy through resorting to a sequence of Semidefinite Programs (SDP). A main drawback of this
approach is that the size of those (SDP) problems may grow intolerably large quickly. Interested
readers may find more information in a recent survey [22] and the references therein.
There is an intrinsic connection between optimizing a polynomial function and the description of all
the polynomial functions that are nonnegative over a given domain. For the case of quadratic poly-
nomials, this connection was explored by Sturm and Zhang in [36], and later for the bi-quadratic
case in Luo et al. [26]. Such investigations can be traced back to the 19th century when the relation-
ship between nonnegative polynomial functions and the sum of squares (SOS) of polynomials was
2
explicitly studied. One concrete question of interest was: Given a multivariate polynomial function
that takes only nonnegative values over the real numbers, can it be represented as a sum of squares
of polynomial functions? Hilbert [19] in 1888 showed that the only three classes of polynomial
functions where this is generically true can be explicitly identified: (1) univariate polynomials; (2)
multivariate quadratic polynomials; (3) bivariate quartic polynomials. Since polynomial functions
with a fixed degree form a vector space, and the nonnegative polynomials and the SOS polynomi-
als form two convex cones respectively within that vector space, the afore-mentioned results can
be understood as a specification of three particular cases where these two convex cones coincide,
while in general of course the cone of nonnegative polynomials is larger. There are certainly other
interesting convex cones in the same vector space. For instance, the convex polynomial functions
form yet another convex cone in that vector space. Helton and Nie [18] introduced the notion of
sos-convex polynomials, to indicate the polynomials whose Hessian matrix can be decomposed as a
sum of squares of polynomial matrices. All these classes of convex cones are important in their own
rights. They are also important for the sake of optimization of polynomial functions. There are
some substantial recent progresses along this direction. As we mentioned earlier, e.g., the question
of Shor [28] regarding the complexity of deciding the convexity of a quartic polynomial was nicely
settled by Ahmadi et al. [2]. It is also natural to inquire if the Hessian matrix of a convex poly-
nomial is sos-convex. Ahmadi and Parrilo [3] gave an example to show that this is not the case in
general. Blekherman proved that a convex polynomial is not necessary a sum of squares [6] if the
degree of the polynomial is larger than two. However, Blekherman’s proof is not constructive, and
it remains an open problem to construct a concrete example of convex polynomial which is not a
sum of squares. Reznick [34] studied the sum of m-th powers of real linear forms.
In view of the cones formed by the polynomial functions (e.g. the cones of nonnegative polynomials,
the convex polynomials, the SOS polynomials and the sos-convex polynomials), it is natural to
inquire about their relational structures. We shall set out to explore this in the paper. In a way there
is a ‘phase transition’ in terms of complexity when the scope of polynomials goes beyond quadratics.
Compared to the quadratic case (cf. Sturm and Zhang [36]), the structure of the quartic forms is far
from being clear. We believe that the class of quartic polynomial functions (or the class of quartic
forms) is an appropriate subject of study on its own right, beyond quadratic functions (or matrices).
There are at least three immediate reasons to elaborate on the quartic polynomials, rather than
polynomial functions of other degrees. First of all, nonnegativity is naturally associated with even
degree polynomials, and the quartic polynomial is next to quadratic polynomials in that hierarchy.
Second, quartic polynomials represent a landscape after the ‘phase transition’ takes place. However,
dealing with quartic polynomials is still manageable, as far as notations are concerned. Finally, from
an application point of view, quartic polynomial optimization is by far the most relevant polynomial
optimization model beyond quadratic polynomials. The afore-mentioned examples such as kurtosis
risks in portfolio management ([20]), the bi-quadratic optimization models ([24]), and the nonlinear
3
least square formulation of sensor network localization ([7]) are all such examples. In this paper
we aim to present the relational structure of several important convex cones in the vector space of
quartic forms. We shall also motivate the study by some examples from applications.
2 Preparations
2.1 Notations
Throughout this paper, we use the lower-case letters to denote vectors (e.g., x ∈ Rn), the capital
letters to denote matrices (e.g., A ∈ Rn2), and the capital calligraphy letters to denote fourth order
tensors (e.g., F ∈ Rn4), with subscriptions of indices being their entries (e.g., x1, Aij ,Fijk` ∈ R).
The boldface capital letters are reserved for sets in the Euclidean space, e.g., various sets of quatric
forms to be introduced later, as well as Rn4, the space of n-dimensional fourth order tensors.
A generic quartic form is a fourth degree homogeneous polynomial function in n variables, or
specifically the function
f(x) =∑
1≤i≤j≤k≤`≤nGijk` xixjxkx`, (1)
where x = (x1, · · · , xn)T ∈ Rn. Closely related to a quartic form is a fourth order super-symmetric
tensor F ∈ Rn4. A tensor is said to be super-symmetric if its entries are invariant under all
permutations of its indices. The set of n-dimensional super-symmetric fourth order tensors is
denoted by Sn4. In fact, super-symmetric tensors are bijectively related to forms. In particular,
restricting to fourth order tensors, for a given super-symmetric tensor F ∈ Sn4, the quartic form
in (1) can be uniquely determined by the following operation:
f(x) = F(x, x, x, x) :=∑
1≤i,j,k,`≤nFijk` xixjxkx`, (2)
where x ∈ Rn, Fijk` = Gijk`/|Π(ijk`)| and Π(ijk`) is the set of all distinctive permutations of the
indices {i, j, k, `}, and vice versa. (This is the same as the one-to-one correspondence between a
symmetric matrix and a quadratic form.) In the remainder of the paper, we shall more frequently
use a super-symmetric tensor F ∈ Sn4
to indicate a quartic form f(x) or F(x, x, x, x), i.e., the
notion of “super-symmetric fourth order tensor” and “quartic form” are used interchangeably in
this paper.
Given a quartic form F ∈ Sn4
and a matrix X ∈ Rn2, we may also define the following operation
(in the same spirit as (2)):
F(X,X) :=∑
1≤i,j,k,`≤nFijk`XijXk`.
4
We call a fourth order tensor G ∈ Rn4partial-symmetric, if
Gijk` = Gjik` = Gij`k = Gk`ij ∀ 1 ≤ i, j, k, ` ≤ n.
Essentially this means that the tensor form is symmetric for the first and the last two indices
respectively, and is also symmetric by swapping the first two and the last two indices. The set of
all partial-symmetric fourth order tensors in Rn4is denoted by
−→S n4
. Obviously Sn4 (−→S n4 ( Rn4
if n ≥ 2.
For any fourth order tensor G ∈ Rn4, we introduce a symmetrization mapping sym : Rn4 7→ Sn
4,
which is F = symG with
Fijk` =1
|Π(ijk`)|∑
π∈Π(ijk`)
Gπ ∀ 1 ≤ i, j, k, ` ≤ n.
The symbol ‘⊗’ represents the outer product of vectors or matrices. If F = x⊗x⊗x⊗x for some x ∈Rn, then Fijk` = xixjxkx`; and if G = X⊗X for some X ∈ Rn2
, then Gijk` = XijXk`. The symbol
‘•’ denotes the operation of inner product. As a result, we have F(x, x, x, x) = F • (x⊗ x⊗ x⊗ x).
2.2 Introducing the Quartic Forms
In this subsection we shall formally introduce the definitions of the quartic forms (or equivalently,
super-symmetric fourth order tensors) discussed in this paper. Let us start with the well known
notion of positive semidefinite (PSD) and the sum of squares (SOS) of polynomials.
Definition 2.1 A quartic form F ∈ Sn4
is called quartic PSD if
F(x, x, x, x) ≥ 0 ∀x ∈ Rn. (3)
The set of all quartic PSD forms in Sn4
is denoted by Sn4
+ .
If a quartic form F ∈ Sn4
can be written as a sum of squares of polynomial functions, then these
polynomials must be quadratic forms, i.e.,
F(x, x, x, x) =
m∑i=1
(xTAix
)2= (x⊗ x⊗ x⊗ x) •
m∑i=1
Ai ⊗Ai,
where Ai ∈ Sn2, the set of symmetric matrices. However,
∑mi=1
(Ai ⊗Ai
)∈−→S n4
is only partial-
symmetric, and may not be exactly F , which must be super-symmetric. To place it in the family
Sn4, a symmetrization operation is required. Since x⊗ x⊗ x⊗ x is super-symmetric, we still have
(x⊗ x⊗ x⊗ x) • sym(∑m
i=1Ai ⊗Ai
)= (x⊗ x⊗ x⊗ x) •
∑mi=1A
i ⊗Ai.
5
Definition 2.2 A quartic form F ∈ Sn4
is called quartic SOS if F(x, x, x, x) is a sum of squares
of quadratic forms, i.e., there exist m symmetric matrices A1, · · · , Am ∈ Sn2
such that
F = sym
(m∑i=1
Ai ⊗Ai)
=m∑i=1
sym(Ai ⊗Ai
).
The set of quartic SOS forms in Sn4
is denoted by Σ2n,4.
As all quartic SOS forms constitute a convex cone, we have
Σ2n,4 = sym cone
{A⊗A |A ∈ Sn
2}.
Usually, for a given F = sym(∑m
i=1Ai ⊗Ai
)it maybe a challenge to write it explicitly as a sum
of squares, although the construction can in principle be done in polynomial-time by semidefinte
programming (SDP), which however is costly. In this sense, having a quartic SOS tensor in super-
symmetric form may not always be beneficial, since the super-symmetry can destroy the SOS
structure.
Since F(X,X) is a quadratic form, the usual sense of nonnegativity carries over. Formally we
introduce this notion below.
Definition 2.3 A quartic form F ∈ Sn4
is called quartic matrix PSD if
F(X,X) ≥ 0 ∀X ∈ Rn2.
The set of quartic matrix PSD forms in Sn4
is denoted by Sn2×n2
+ .
Related to the sum of squares for quartic forms, we now introduce the notion to the sum of quartics
(SOQ): If a quartic form F ∈ Sn4
is SOQ, then there are m vectors a1, · · · , am ∈ Rn such that
F(x, x, x, x) =
m∑i=1
(xTai
)4= (x⊗ x⊗ x⊗ x) •
m∑i=1
ai ⊗ ai ⊗ ai ⊗ ai.
Definition 2.4 A quartic form F ∈ Sn4
is called quartic SOQ if F(x, x, x, x) is a sum of fourth
powers of linear forms, i.e., there exist m vectors a1, · · · , am ∈ Rn such that
F =
m∑i=1
ai ⊗ ai ⊗ ai ⊗ ai.
The set of quartic SOQ forms in Sn4
is denoted by Σ4n,4.
6
As all quartic SOQ forms also constitute a convex cone, we denote
Σ4n,4 = cone {a⊗ a⊗ a⊗ a | a ∈ Rn} ⊆ Σ2
n,4.
In the case of quadratic functions, it is well known that for a given homogeneous form (i.e., a
symmetric matrix, for that matter) A ∈ Sn2
the following two statements are equivalent:
• A is positive semidefinite (PSD): A(x, x) := xTAx ≥ 0 for all x ∈ Rn.
• A is a sum of squares (SOS): A(x, x) =∑m
i=1(xTai)2 (or equivalently A =∑m
i=1 ai ⊗ ai) for
some a1, · · · , am ∈ Rn.
It is therefore clear that the four types of quartic forms defined above are actually different exten-
sions of the nonnegativity. In particular, quartic PSD and quartic matrix PSD forms are extended
from quadratic PSD, while quartic SOS and SOQ forms are in the form of summation of non-
negative polynomials, and are extended from quadratic SOS. We will show later that there is an
interesting hierarchical relationship:
Σ4n,4 ( Sn
2×n2
+ ( Σ2n,4 ( Sn
4
+ . (4)
Recently, a class of polynomials termed the sos-convex polynomials (cf. Helton and Nie [18]) has
been brought to attention, which is defined as follows (see [4] for three other equivalent definitions
of the sos-convexity):
A multivariate polynomial function f(x) is sos-convex if its Hessian matrix H(x) can
be factorized as H(x) = (M(x))TM(x) with a polynomial matrix M(x).
The reader is referred to [3] for applications of the sos-convex polynomials. In this paper, we shall
focus on Sn4
and investigate sos-convex quartic forms with the hierarchy (4). For a quartic form
F ∈ Sn4, it is straightforward to compute its Hessian matrix H(x) = 12F(x, x, ·, ·), i.e.,
(H(x))ij = 12F(x, x, ei, ej) ∀ 1 ≤ i, j ≤ n,
where ei ∈ Rn is the vector whose i-th entry is 1 and other entries are zeros. Therefore H(x) is a
quadratic matrix of x. If H(x) can be decomposed as H(x) = (M(x))TM(x) with M(x) being a
polynomial matrix, then M(x) must be linear with respect to x.
Definition 2.5 A quartic form F ∈ Sn4
is called quartic sos-convex, if there exists a linear matrix
M(x) of x, such that its Hessian matrix
12F(x, x, ·, ·) = (M(x))TM(x).
The set of quartic sos-convex forms in Sn4
is denoted by Σ2∇2
n,4.
7
Helton and Nie [18] proved that a nonnegative polynomial is sos-convex, then it must be SOS.
In particular, if the polynomial is a quartic form, by denoting the i-th row of the linear matrix
M(x) to be xTAi for i = 1, . . . ,m and some matrices A1, · · · , Am ∈ Rn2, then (M(x))TM(x) =∑m
i=1(Ai)TxxTAi. Therefore
F(x, x, x, x) = xTF(x, x, ·, ·)x =1
12xT(M(x))TM(x)x =
1
12
m∑i=1
(xTAix
)2 ∈ Σ2n,4.
In addition, the Hessian matrix for a quartic sos-convex form is obviously positive semidefinite for
any x ∈ Rn. Hence sos-convexity implies convexity. Combining these two facts, we conclude that
a quartic sos-convex form is both SOS and convex, which motivates us to study the last quartic
forms in this paper.
Definition 2.6 A quartic form F ∈ Sn4
is called convex and SOS, if it is both quartic SOS and
convex. The set of quartic convex and SOS forms in Sn4
is denoted by Σ2n,4
⋂Sn
4
cvx.
Here Sn4
cvx is denoted to be the set of all convex quartic forms in Sn4.
2.3 Main Results and the Organization of the Paper
All the sets of the quartic forms defined in Section 2.2 are clearly convex cones. The remainder of
this paper is organized as follows. In Section 3, we start by studying the cones: Sn4
+ , Σ2n,4, Sn
2×n2
+ ,
and Σ4n,4. First of all, we show that they are all closed, and that they can be presented in different
formulations. As an example, the cone of quartic SOQ forms is
Σ4n,4 = cone {a⊗ a⊗ a⊗ a | a ∈ Rn} = sym cone
{A⊗A |A ∈ Sn
2
+ , rank(A) = 1},
which can also be written as
sym cone{A⊗A |A ∈ Sn
2
+
},
meaning that the rank-one constraint can be removed without affecting the cone itself. Moreover,
among these four cones there are two primal-dual pairs: Sn4
+ is dual to Σ4n,4, and Σ2
n,4 is dual to
Sn2×n2
+ . The hierarchical relationship Σ4n,4 ( Sn
2×n2
+ ( Σ2n,4 ( Sn
4
+ will also be shown in the same
section.
In Section 4, we further study two more cones: Σ2∇2
n,4and Σ2
n,4
⋂Sn
4
cvx. Interestingly, these two new
cones can be nicely placed in the hierarchical scheme (4):
Σ4n,4 ( Sn
2×n2
+ ( Σ2∇2
n,4⊆(Σ2n,4
⋂Sn
4
cvx
)( Σ2
n,4 ( Sn4
+ . (5)
The complexity status of all these cones are discussed in Section 5, and in particular we show that
Σ2∇2
n,4((Σ2n,4
⋂Sn
4
cvx
)unless P = NP , completing the picture presented in (5), on the premise
8
that P 6= NP . The low dimensional cases of these cones are also discussed in Section 5. Specially,
for the case n = 2, all the six cones reduce to only two distinctive ones, and for the case n = 3,
they reduce to exactly three distinctive cones. Finally, in Section 6 we discuss applications of
quartic conic programming, including bi-quadratic assignment problems and eigenvalues of super-
symmetric tensors.
3 Quartic PSD Forms, Quartic SOS Forms, and the Dual Cones
Let us now consider the first four cones of quartic forms introduced in Section 2.2: Σ4n,4, Sn
2×n2
+ ,
Σ2n,4, and Sn
4
+ .
3.1 Closedness
Proposition 3.1 Σ4n,4, Sn
2×n2
+ , Σ2n,4, and Sn
4
+ are all closed convex cones.
While Sn4
+ and Sn2×n2
+ are evidently closed, by a similar argument as in [36] it is also easy to see
that the cone of quartic SOS forms Σ2n,4 := sym cone
{A⊗A |A ∈ Sn
2}
is closed. It only remains
to show that the cone of quartic SOQ forms Σ4n,4 is closed. In fact, we have a slightly stronger
result below, which ensures the closedness of Σ4n,4:
Lemma 3.2 If D ⊆ Rn, then
cl cone {a⊗ a⊗ a⊗ a | a ∈ D} = cone {a⊗ a⊗ a⊗ a | a ∈ cl D} .
Proof. Suppose that F ∈ cl cone {a⊗ a⊗ a⊗ a | a ∈ D}, then there is a sequence of quartic forms
Fk ∈ cone {a⊗ a⊗ a⊗ a | a ∈ D} (k = 1, 2, . . .), such that F = limk→∞Fk. Since the dimension
of Sn4
is N :=(n+3
4
), it follows from Caratheodory’s theorem that for any given Fk, there exists
an n× (N + 1) matrix Zk, such that
Fk =
N+1∑i=1
zk(i)⊗ zk(i)⊗ zk(i)⊗ zk(i),
where zk(i) is the i-th column vector of Zk, and is a positive multiple of a vector in D. Now define
trFk =∑n
j=1Fkjjjj , thenN+1∑i=1
n∑j=1
(Zkji)4 = trFk → trF .
9
Thus, the sequence {Zk} is bounded, and have a cluster point Z∗, satisfying F =∑N+1
i=1 z∗(i) ⊗z∗(i)⊗ z∗(i)⊗ z∗(i). Note that each column of Z∗ is also a positive multiple of a vector in cl D, it
follows that F ∈ cone {a⊗ a⊗ a⊗ a | a ∈ cl D}. The converse containing relationship is trivial. �
The cone of quartic SOQ forms is closely related to the fourth moment of a multi-dimensional
random variable. Given an n-dimensional random variable ξ = (ξ1, · · · , ξn)T on the support set
D ⊆ Rn with density function p, its fourth moment is a super-symmetric fourth order tensor
M∈ Sn4, whose (i, j, k, `)-th entry is
Mijk` = E [ξiξjξkξ`] =
∫Dxixjxkx` p(x)dx.
Suppose the fourth moment of ξ is finite, then by the closedness of Σ4n,4, we have
M = E [ξ ⊗ ξ ⊗ ξ ⊗ ξ] =
∫D
(x⊗ x⊗ x⊗ x) p(x)dx ∈ cone {a⊗ a⊗ a⊗ a | a ∈ Rn} = Σ4n,4.
Conversely, for anyM∈ Σ4n,4, it is easy to verify that there exists an n-dimensional random variable
whose fourth moment is just M. Thus, the set of all the finite fourth moments of n-dimensional
random variables is exactly Σ4n,4, similar to the fact that all possible covariance matrices form the
cone of positive semidefinite matrices.
3.2 Alternative Representations
In this subsection we present some alternative forms of the same cones that we have discussed.
Some of these alternative representations are more convenient to use in various applications.
Theorem 3.3 For the quartic polynomials cones introduced, we have the following equivalent rep-
resentations:
1. For the cone of quartic SOS forms
Σ2n,4 := sym cone
{A⊗A |A ∈ Sn
2}
= sym{F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Sn
2}
;
2. For the cone of quartic matrix PSD forms
Sn2×n2
+ :={F ∈ Sn
4 | F(X,X) ≥ 0, ∀X ∈ Rn2}
={F ∈ Sn
4 | F(X,X) ≥ 0, ∀X ∈ Sn2}
= Sn4⋂
cone{A⊗A |A ∈ Sn
2}
;
10
3. For the cone of quartic SOQ forms
Σ4n,4 := cone {a⊗ a⊗ a⊗ a | a ∈ Rn} = sym cone
{A⊗A |A ∈ Sn
2
+
}.
Recall that−→S n4
is the set of partial-symmetric fourth order tensors in Rn4, defined in Section 2.1.
The remaining of this subsection is devoted to the proof of Theorem 3.3.
Let us first study the equivalent representations for Σ2n,4 and Sn
2×n2
+ . To verify a quartic matrix
PSD form, we should check the operations of quartic forms on matrices. In fact, the quartic matrix
PSD forms can be extended to the space of partial-symmetric tensors−→S n4
. It is not hard to verify
that for any F ∈−→S n4
, it holds that
F(X,Y ) = F(XT, Y ) = F(X,Y T) = F(Y,X) ∀X,Y ∈ Rn2, (6)
which implies that F(X,Y ) is invariant under the transpose operation as well as the operation to
swap the X and Y matrices. Indeed, it is easy to see that the partial-symmetry of F is a necessary
and sufficient condition for (6) to hold. We have the following property for quartic matrix PSD
forms in−→S n4
, similar to that for Sn2×n2
+ in Theorem 3.3.
Lemma 3.4 For partial-symmetric four order tensors, it holds that
−→S n2×n2
+ :={F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Rn2
}=
{F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Sn
2}
(7)
= cone{A⊗A |A ∈ Sn
2}. (8)
Proof. Observe that for any skew-symmetric Y ∈ Rn2, i.e., Y T = −Y , we have
F(X,Y ) = −F(X,−Y ) = −F(X,Y T) = −F(X,Y ) ∀X ∈ Rn2,
which implies that F(X,Y ) = 0. As any square matrix can be written as the sum of a symmetric
matrix and a skew-symmetric matrix, say for Z ∈ Rn2, by letting X = (Z + ZT)/2 which is
symmetric, and Y = (Z − ZT)/2 which is skew-symmetric, we have Z = X + Y . Therefore,
F(Z,Z) = F(X + Y,X + Y ) = F(X,X) + 2F(X,Y ) + F(Y, Y ) = F(X,X).
This implies the equivalence between F(X,X) ≥ 0, ∀X ∈ Rn2and F(X,X) ≥ 0, ∀X ∈ Sn
2, which
proves (7).
To prove (8), first note that cone{A⊗A |A ∈ Sn
2}⊆ {F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Rn2}.
Conversely, given any G ∈−→S n4
with G(X,X) ≥ 0, ∀X ∈ Rn2, we may rewrite G as an n2 × n2
symmetric matrix MG . Therefore
(vec (X))TMG vec (X) = G(X,X) ≥ 0 ∀X ∈ Rn2,
11
which implies that MG is positive semidefinite. Let MG =∑m
i=1 zi(zi)T, where
zi =(zi11, · · · , zi1n, · · · , zin1, · · · , zinn
)T ∀ 1 ≤ i ≤ m.
Note that for any 1 ≤ k, ` ≤ n, Gk``k =∑m
i=1 zik`z
i`k, Gk`k` =
∑mi=1(zik`)
2 and G`k`k =∑m
i=1(zi`k)2,
as well as Gk``k = Gk`k` = G`k`k by partial-symmetry of G. We have
m∑i=1
(zik` − zi`k)2 =
m∑i=1
(zik`)2 +
m∑i=1
(zi`k)2 − 2
m∑i=1
zik`zi`k = Gk`k` + G`k`k − 2Gk``k = 0,
which implies that zik` = zi`k for any 1 ≤ k, ` ≤ n. Therefore, we may construct a symmetric matrix
Zi ∈ Sn2, such that vec (Zi) = zi for all 1 ≤ i ≤ m. We have G =
∑mi=1 Z
i ⊗ Zi, and so (8) is
proven. �
The first part of Theorem 3.3 follows from (8) by applying the symmetrization operation on both
sides, while the second part of Theorem 3.3 follows from (7) and (8) by restricting to Sn4.
Let us now turn to proving the last part of Theorem 3.3, which is an alternative representation of
the quartic SOQ forms. Obviously we need only to show that
sym cone{A⊗A |A ∈ Sn
2
+
}⊆ cone {a⊗ a⊗ a⊗ a | a ∈ Rn} .
Since there is a one-to-one mapping from quartic forms to fourth order super-symmetric tensors, it
suffices to show that for any A ∈ Sn2
+ , the function (xTAx)2 can be written as a form of∑m
i=1(xTai)4
for some a1, · · · , am ∈ Rn. Note that the so-called Hilbert’s identity (see e.g., Barvinok [5]) asserts
the following:
For any fixed positive integers d and n, there always existm real vectors a1, · · · , am ∈ Rn
such that (xTx)d =∑m
i=1(xTai)2d.
In fact, when d = 2, He et al. [17] proposed a polynomial-time algorithm to find the afore-mentioned
representations where the number m is bounded by a polynomial of n, although in the original
version of Hilbert’s identity m is exponential in n. Since we have A ∈ Sn2
+ , replacing x by A1/2y in
Hilbert’s identity when d = 2, one gets (yTAy)2 =∑m
i=1(yTA1/2ai)4. The desired decomposition
follows, and this proves the last part of Theorem 3.3.
3.3 Duality
In this subsection, we shall discuss the duality relationships among these four cones of quartic forms.
Note that Sn4
is the ground vector space within which the duality is defined, unless otherwise
specified.
12
Theorem 3.5 The cone of quartic PSD forms and the cone of quartic SOQ forms are primal-dual
pair, i.e., Σ4n,4 =
(Sn
4
+
)∗and Sn
4
+ =(Σ4n,4
)∗. The cone of quartic SOS forms and the cone of
quartic matrix PSD forms are primal-dual pair, i.e., Sn2×n2
+ =(Σ2n,4
)∗and Σ2
n,4 =(Sn
2×n2
+
)∗.
Remark that the primal-dual relationship between Σ4n,4 and Sn
4
+ was also studied in [34]. Let us
indeed start by discussing the primal-dual pair Σ4n,4 and Sn
4
+ . In Proposition 1 of [36], Sturm and
Zhang proved that for the quadratic forms, {A ∈ Sn2 |xTAx ≥ 0, ∀x ∈ D} and cone {aaT | a ∈ D}
are a primal-dual pair for any closed D ⊆ Rn. We observe that a similar structure holds for the
quartic forms as well. The first part of Theorem 3.5 then follows from next lemma.
Lemma 3.6 If D ⊆ Rn is closed, then Sn4
+ (D) := {F ∈ Sn4 | F(x, x, x, x) ≥ 0, ∀x ∈ D} and
cone {a⊗ a⊗ a⊗ a | a ∈ D} are a primal-dual pair, i.e.,
Sn4
+ (D) = (cone {a⊗ a⊗ a⊗ a | a ∈ D})∗ (9)
and (Sn
4
+ (D))∗
= cone {a⊗ a⊗ a⊗ a | a ∈ D}.
Proof. Since cone {a⊗a⊗a⊗a | a ∈ D} is closed by Lemma 3.2, we only need to show (9). In fact,
if F ∈ Sn4
+ (D), then F • (a ⊗ a ⊗ a ⊗ a) = F(a, a, a, a) ≥ 0 for all a ∈ D. Thus F • G ≥ 0 for all
G ∈ cone {a⊗a⊗a⊗a | a ∈ D}, which implies that F ∈ (cone {a⊗ a⊗ a⊗ a | a ∈ D})∗. Conversely,
if F ∈ (cone {a⊗ a⊗ a⊗ a | a ∈ D})∗, then F • G ≥ 0 for all G ∈ cone {a ⊗ a ⊗ a ⊗ a | a ∈ D}. In
particular, by letting G = x ⊗ x ⊗ x ⊗ x, we have F(x, x, x, x) = F • (x ⊗ x ⊗ x ⊗ x) ≥ 0 for all
x ∈ D, which implies that F ∈ Sn4
+ (D). �
Let us turn to the primal-dual pair of Sn2×n2
+ and Σ2n,4. For technical reasons, we shall momentarily
lift the ground space from Sn4
to the space of partial-symmetric tensors−→S n4
. This enlarges all
the dual objects. To distinguish these two dual objects, let us use the notation ‘K−→∗ ’ to indicate
the dual of convex cone K ∈ Sn4 ⊆−→S n4
generated in the space−→S n4
, while ‘K∗’ is the dual of K
generated in the space Sn4.
Lemma 3.7 For partial-symmetric tensors, the cone−→S n2×n2
+ is self-dual with respect to the space−→S n4
, i.e.,−→S n2×n2
+ =(−→
S n2×n2
+
)−→∗.
Proof. According to Proposition 1 of [36] and the partial-symmetry of−→S n4
, we have(cone
{A⊗A |A ∈ Sn
2})−→∗
={F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Sn
2}.
13
By Lemma 3.4, we have
−→S n2×n2
+ ={F ∈
−→S n4 | F(X,X) ≥ 0, ∀X ∈ Sn
2}
= cone{A⊗A |A ∈ Sn
2}.
Thus−→S n2×n2
+ is self-dual with respect to the space−→S n4
. �
Notice that by definition and Lemma 3.7, we have
Σ2n,4 = sym cone
{A⊗A |A ∈ Sn
2}
= sym−→S n2×n2
+ = sym(−→
S n2×n2
+
)−→∗,
and by the alternative representation in Theorem 3.3 we have
Sn2×n2
+ = Sn4⋂
cone{A⊗A |A ∈ Sn
2}
= Sn4⋂−→
S n2×n2
+ .
Therefore the duality between Sn2×n2
+ and Σ2n,4 follows immediately from the following lemma.
Lemma 3.8 If K ⊆−→S n4
is a closed convex cone and K−→∗ is its dual with respect to the space
−→S n4
,
then K⋂
Sn4
and sym K−→∗ are a primal-dual pair with respect to the space Sn
4, i.e.,
(K⋂
Sn4)∗
=
sym K−→∗ and K
⋂Sn
4=(
sym K−→∗)∗
.
Proof. For any G ∈ sym K−→∗ ⊆ Sn
4, there is a G′ ∈ K
−→∗ ⊆−→S n4
, such that G = symG′ ∈ Sn4. We
then have Gijk` = 13(G′ijk` + G′ikj` + G′i`jk). Thus for any F ∈ K
⋂Sn
4 ⊆ Sn4, it follows that
F • G =∑
1≤i,j,k,`≤n
Fijk`(G′ijk` + G′ikj` + G′i`jk)3
=∑
1≤i,j,k,`≤n
Fijk`G′ijk` + Fikj`G′ikj` + Fi`jkG′i`jk3
= F • G′ ≥ 0.
Therefore G ∈(K⋂
Sn4)∗
, implying that sym K−→∗ ⊆
(K⋂
Sn4)∗
.
Moreover, if F ∈(
sym K−→∗)∗⊆ Sn
4, then for any G′ ∈ K
−→∗ ⊆−→S n4
, we have G = symG′ ∈ sym K−→∗ ,
and G′ • F = G • F ≥ 0. Therefore F ∈(K−→∗)−→∗
= cl K = K, which implies that(
sym K−→∗)∗⊆(
K⋂
Sn4)
. Finally, the duality relationship holds by the bipolar theorem and the closedness of
these cones. �
3.4 The Hierarchical Structure
The last part of this section is to present a hierarchy among these four cones of quartic forms. The
main result is summarized in the theorem below.
14
Theorem 3.9 If n ≥ 4, then
Σ4n,4 ( Sn
2×n2
+ ( Σ2n,4 ( Sn
4
+ .
For the low dimension cases (n ≤ 3), we shall present it in Section 5.2. Evidently a quartic
SOS form is quartic PSD, implying Σ2n,4 ⊆ Sn
4
+ . By invoking the duality operation and using
Theorem 3.5 we have Σ4n,4 ⊆ Sn
2×n2
+ , while by the alternative representation in Theorem 3.3
we have Sn2×n2
+ = Sn4 ⋂
cone{A⊗A |A ∈ Sn
2}
, and by the very definition we have Σ2n,4 =
sym cone{A⊗A |A ∈ Sn
2}
. Therefore Sn2×n2
+ ⊆ Σ2n,4. Finally, the strict containing relationship
is a result of the following examples.
Example 3.1 (Quartic forms in Sn4
+ \Σ2n,4 when n = 4). Let g1(x) = x1
2(x1−x4)2+x22(x2−x4)2+
x32(x3−x4)2 +2x1x2x3(x1 +x2 +x3−2x4) and g2(x) = x1
2x22 +x2
2x32 +x3
2x12 +x4
4−4x1x2x3x4,
then both g1(x) and g2(x) are in S44+ \ Σ2
4,4. We refer the interested readers to [33] for more
information on these two cones.
Example 3.2 (A quartic form in Sn2×n2
+ \ Σ4n,4 when n = 4). Construct F ∈ S44, whose only
nonzero entries (taking into account the super-symmetry) are F1122 = 4, F1133 = 4, F2233 = 4,
F1144 = 9, F2244 = 9, F3344 = 9, F1234 = 6, F1111 = 29, F2222 = 29, F3333 = 29, and F4444 =
3 + 257 . One may verify straightforwardly that F can be decomposed as
∑7i=1A
i ⊗ Ai, with A1 =√
7 0 0 0
0√
7 0 0
0 0√
7 0
0 0 0 5√7
, A2 =
0 2 0 0
2 0 0 0
0 0 0 3
0 0 3 0
, A3 =
0 0 2 0
0 0 0 3
2 0 0 0
0 3 0 0
, A4 =
0 0 0 3
0 0 2 0
0 2 0 0
3 0 0 0
,
A5 =
−2 0 0 0
0 3 0 0
0 0 3 0
0 0 0 1
, A6 =
3 0 0 0
0 −2 0 0
0 0 3 0
0 0 0 1
, and A7 =
3 0 0 0
0 3 0 0
0 0 −2 0
0 0 0 1
. According to
Theorem 3.3, we have F ∈ S42×42
+ . Recall g2(x) in Example 3.1, which is quartic PSD. Denote Gto be the super-symmetric tensor associated with g2(x), thus G ∈ S44
+ . One computes that G • F =
4 + 4 + 4 + 3 + 257 − 24 < 0. By the duality result as stipulated in Theorem 3.5, we conclude that
F /∈ Σ44,4.
Example 3.3 (A quartic form in Σ2n,4 \ Sn
2×n2
+ when n = 3). Let g3(x) = 2x14 + 2x2
4 + 12x3
4 +
6x12x3
2 + 6x22x3
2 + 6x12x2
2, which is obviously quartic SOS. Now recycle the notation and denote
G ∈ Σ23,4 to be the super-symmetric tensor associated with g3(x), and we have G1111 = 2, G2222 = 2,
15
G3333 = 12 , G1122 = 1, G1133 = 1, and G2233 = 1. If we let X = Diag (1, 1,−4) ∈ S32, then
G(X,X) =∑
1≤i,j,k,`≤3
Gijk`XijXk` =∑
1≤i,k≤3
GiikkXiiXkk =
1
1
−4
T 2 1 1
1 2 1
1 1 12
1
1
−4
= −2,
implying that G /∈ S32×32
+ .
4 Cones Related to Convex Quartic Forms
In this section we shall study the cone of quartic sos-convex forms Σ2∇2
n,4, and the cone of quartic
forms which are both SOS and convex Σ2n,4
⋂Sn
4
cvx. The aim is to incorporate these two new cones
into the hierarchical structure as depicted in Theorem 3.9.
Theorem 4.1 If n ≥ 4, then
Σ4n,4 ( Sn
2×n2
+ ( Σ2∇2
n,4⊆(Σ2n,4
⋂Sn
4
cvx
)( Σ2
n,4 ( Sn4
+ .
As mentioned in Section 2.2, an sos-convex homogeneous quartic polynomial function is both SOS
and convex (see also [3]), which implies that Σ2∇2
n,4⊆(Σ2n,4
⋂Sn
4
cvx
). Moreover, the following
example shows that a quartic SOS form is not necessarily convex.
Example 4.1 (A quartic form in Σ2n,4 \ Sn
4
cvx when n = 2). Let g4(x) = (xTAx)2 with A ∈ Sn2,
and its Hessian matrix is ∇2g4(x) = 8AxxTA+ 4xTAxA. In particular, by letting A =
[−3 0
0 1
]
and x =
(0
1
), we have ∇2f(x) =
[0 0
0 8
]+
[−12 0
0 4
]� 0, implying that g4(x) is not convex.
The above example suggests that(Σ2n,4
⋂Sn
4
cvx
)( Σ2
n,4 when n ≥ 2. Next we shall prove the
assertion that Sn2×n2
+ ( Σ2∇2
n,4when n ≥ 3. To this end, let us first quote a result on the sos-convex
functions due to Ahmadi and Parrilo [3]:
If f(x) is a polynomial with its Hessian matrix being ∇2f(x), then f(x) is sos-convex
if and only if yT∇2f(x)y is a sum of squares in (x, y).
For a quartic form F(x, x, x, x), its Hessian matrix is 12F(x, x, ·, ·). Therefore, F is quartic sos-
convex if and only if F(x, x, y, y) is is a sum of squares in (x, y). Now if F ∈ Sn2×n2
+ , then by
16
Theorem 3.3 we may find matrices A1, · · · , Am ∈ Sn2
such that F =∑m
t=1At ⊗At. We have
F(x, x, y, y) = F(x, y, x, y) =m∑t=1
∑1≤i,j,k,`≤n
xiyjxky`AtijA
tk`
=m∑t=1
∑1≤i,j≤n
xiyjAtij
∑1≤k,`≤n
xky`Atk`
=m∑t=1
(xTAty
)2,
which is a sum of squares in (x, y), hence sos-convex. This proves Sn2×n2
+ ⊆ Σ2∇2
n,4, and the example
below rules out the equality when n ≥ 3.
Example 4.2 (A quartic form in Σ2∇2
n,4\ Sn
2×n2
+ when n = 3). Recall g3(x) = 2x14 + 2x2
4 +
12x3
4 + 6x12x3
2 + 6x22x3
2 + 6x12x2
2 in Example 3.3, which is shown not to be quartic matrix PSD.
Moreover, it is straightforward to compute that
∇2g3(x) = 24
x1
x2
x32
x1
x2
x32
T
+ 12
0
x3
x2
0
x3
x2
T
+ 12
x3
0
x1
x3
0
x1
T
+ 12
x22 0 0
0 x12 0
0 0 0
� 0,
which implies that g3(x) is quartic sos-convex.
A natural question regarding to the hierarchical structure in Theorem 4.1 is whether Σ2n,4
⋂Sn
4
cvx =
Σ2∇2
n,4or not. In fact, the relationship between convex, sos-convex, and SOS is a highly interesting
subject which attracted many speculations in the literature. Ahmadi and Parrilo [3] gave an
explicit example with a three dimensional homogeneous form of degree eight, which they showed
to be both convex and SOS but not sos-convex. However, for quartic polynomials (degree four),
such an explicit instant in(Σ2n,4
⋂Sn
4
cvx
)\Σ2
∇2n,4
is not in sight. Notwithstanding the difficulty in
constructing an explicit quartic example, on the premise that P 6= NP in Section 5 we will show
that(Σ2n,4
⋂Sn
4
cvx
)6= Σ2
∇2n,4
. With that piece of information, the chain of containing relations
manifested in Theorem 4.1 is complete, under the assumption that P 6= NP . However, the following
open question remains:
Question 4.1 Find an explicit instant in(Σ2n,4
⋂Sn
4
cvx
)\Σ2∇2
n,4: a quartic form that is both SOS
and convex, but not sos-convex.
The two newly introduced cones in this section are related to the convexity properties. Some more
words on convex quartic forms are in order here. As mentioned in Section 2.2, for a quartic form
F ∈ Sn4
cvx, its Hessian matrix is 12F(x, x, ·, ·). Therefore, F is convex if and only if F(x, x, ·, ·) � 0
for all x ∈ Rn, which is equivalent to F(x, x, y, y) ≥ 0 for all x, y ∈ Rn. In fact, it is also equivalent
17
to F(X,Y ) ≥ 0 for all X,Y ∈ Sn2
+ . To see why, we first decompose the positive semidefinite matrices
X and Y , and let X =∑n
i=1 xi(xi)T and Y =
∑nj=1 y
j(yj)T (see e.g., Sturm and Zhang [36]). Then
F(X,Y ) = F
n∑i=1
xi(xi)T,
n∑j=1
yj(yj)T
=
∑1≤i,j≤n
F(xi(xi)T, yj(yj)T
)=
∑1≤i,j≤n
F(xi, xi, yj , yj
)≥ 0,
if F(x, x, y, y) ≥ 0 for all x, y ∈ Rn. Note that the converse is trivial, as it reduces to let X and Y
be rank-one positive semidefinite matrices. Thus we have the following equivalence for the quartic
convex forms.
Proposition 4.2 For a given quartic form F ∈ Sn4, the following statements are equivalent:
• F(x, x, x, x) is convex;
• F(x, x, ·, ·) is positive semidefinite for all x ∈ Rn;
• F(x, x, y, y) ≥ 0 for all x, y ∈ Rn;
• F(X,Y ) ≥ 0 for all X,Y ∈ Sn2
+ .
For the relationship between the cone of quartic convex forms and the cone of quartic SOS forms,
Example 4.1 has ruled out the possibility that Σ2n,4 ⊆ Sn
4
cvx, while Blekherman [6] proved that
Σ2n,4 is not contained in Sn
4
cvx. Therefore these two cones are indeed distinctive. According to
Blekherman [6], the cone of quartic convex forms is actually much bigger than the cone of quartic
SOS forms when n is sufficiently large. However, at this point we are not aware of any explicit
instant belonging to Sn4
cvx \Σ2n,4. In fact, according to a recent working paper of Ahmadi et al. [1],
this kind of instants exist only when n ≥ 4. Anyway, the following challenge remains:
Question 4.2 Find an explicit instant in Sn4
cvx \ Σ2n,4: a quartic convex form that is not quartic
SOS.
5 Complexity and Low Dimensions of the Quartic Cones
In this section, we shall study the computational complexity issues for the membership queries
regarding these cones of quartic forms. Unlike their quadratic counterparts where the positive
18
semidefniteness can be checked in polynomial-time, the case for the quartic cones are substantially
subtler. We also study the low dimension cases of these cones, as a complement to the result of
hirarchary relationship in Theorem 4.1.
5.1 Complexity
Let us start with easy cases. It is well known that deciding whether a general polynomial function is
SOS can be done in polynomial-time, by resorting to checking the feasibility of an SDP. Therefore,
the membership query for Σ2n,4 can be done in polynomial-time. By the duality relationship claimed
in Theorem 3.5, the membership query for Sn2×n2
+ can also be done in polynomial-time. In fact,
for any quartic form F ∈ Sn4, we may rewrite F as an n2 × n2 matrix, to be denoted by MF , and
then Theorem 3.3 assures that F ∈ Sn2×n2
+ if and only if MF is positive semedefinite, which can be
checked in polynomial-time. Moreover, as discussed in Section 4, a quartic form F is sos-convex if
and only if yT(∇2F(x, x, x, x)
)y = 12F(x, x, y, y) is a sum of squares in (x, y), which can also be
checked in polynomial-time. Therefore, the membership checking problem for Σ2∇2
n,4can be done
in polynomial-time as well. Summarizing, we have:
Proposition 5.1 Whether a quartic form belongs to Σ2n,4, Sn
2×n2
+ , or Σ2∇2
n,4, can be verified in
polynomial-time.
Unfortunately, the membership checking problems for all the other cones that we have discussed so
far are difficult. To see why, let us introduce a famous cone of quadratic functions: the co-positive
cone
C :={A ∈ Sn
2 |xTAx ≥ 0, ∀x ∈ Rn+
},
whose membership query is known to be co-NP-complete. The dual of the co-positive cone is the
cone of completely positive matrices, defined as
C∗ := cone{xxT |x ∈ Rn
+
}.
Recently, Dickinson and Gijben [14] gave a formal proof for the NP-completeness of the membership
problem for C∗.
Proposition 5.2 It is co-NP-complete to check if a quartic form belongs to Sn4
+ (the cone of quartic
PSD forms).
Proof. We shall reduce the problem to checking the membership of the co-positive cone C. In
particular, given any matrix A ∈ Sn2, construct an F ∈ Sn
4, whose only nonzero entries are
Fiikk = Fikik = Fikki = Fkiik = Fkiki = Fkkii =
{Aik3 i 6= k,
Aik i = k.(10)
19
For any x ∈ Rn,
F(x, x, x, x) =∑
1≤i<k≤n(Fiikk + Fikik + Fikki + Fkiik + Fkiki + Fkkii)xi2xk2 +
n∑i=1
Fiiiixi4
=∑
1≤i,k≤nAikxi
2xk2 = (x ◦ x)TA(x ◦ x), (11)
where the symbol ‘◦’ represents the Hadamard product. Denote y = x ◦ x ≥ 0, and then
F(x, x, x, x) ≥ 0 if and only if yTAy ≥ 0. Therefore A ∈ C if and only if F ∈ Sn4
+ and the
reduction is complete. �
Proposition 5.3 It is NP-hard to check if a quartic form belongs to Σ4n,4 (the cone of quartic SOQ
forms).
Proof. Similarly, the problem can be reduced to checking the membership of the completely positive
cone C∗. In particular, given any matrix A ∈ Sn2, construct an F ∈ Sn
4, whose only nonzero entries
are defined exactly as in (10). If A ∈ C∗, then A =∑m
t=1 at(at)T for some a1, · · · , am ∈ Rn
+. By
the construction of F , we have
Fiikk = Fikik = Fikki = Fkiik = Fkiki = Fkkii =
{ ∑mt=1
atiatk
3 i 6= k,∑mt=1
(ati)2
i = k.
Denote At = Diag (at) ∈ Sn2
+ for all 1 ≤ t ≤ m. It is straightforward to verify that
F =m∑t=1
sym(At ⊗At
)= sym
(m∑t=1
At ⊗At).
Therefore by Theorem 3.3 we have F ∈ Σ4n,4.
Conversely, if A /∈ C∗, then there exits a vector y ∈ Rn+, such that yTAy < 0. Define a vector
x ∈ Rn+ with xi =
√yi for all 1 ≤ i ≤ n. By (11), we have
F • (x⊗ x⊗ x⊗ x) = F(x, x, x, x) = (x ◦ x)TA(x ◦ x) = yTAy < 0.
Therefore, by the duality relationship in Theorem 3.5, we have F /∈ Σ4n,4. Since A ∈ C∗ if and only
if F ∈ Σ4n,4 and so it follows that Σ4
n,4 is a hard cone. �
Recently, Burer [8] showed that a large class of mixed-binary quadratic programs can be formulated
as co-positive programs where a linear function is minimized over a linearly constrained subset of
the cone of completely positive matrices. Later, Burer and Dong [9] extended this equivalence to
general nonconvex quadratically constrained quadratic program whose feasible region is nonempty
20
and bounded. From the proof of Proposition 5.3, the cone of completely positive matrices can be
imbedded into the cone of quartic SOQ forms. Evidently, these mixed-binary quadratic programs
can also be formulated as linear conic program with the cone Σ4n,4. In fact, the modeling power of
Σ4n,4 is much greater, which we shall discuss in Section 6 for further illustration.
Before concluding this subsection, a final remark on the cone Σ2n,4
⋂Sn
4
cvx is in order. Recall the
recent breakthrough [2] mentioned in Section 1, that checking the convexity of a quartic form
is strongly NP-hard. However, if we are given more information, that the quartic form to be
considered is a sum of squares, will this make the membership easier? The answer is still no, as
the following theorem asserts.
Theorem 5.4 Deciding the convexity of a quartic SOS form is strongly NP-hard. In particular, it
is strongly NP-hard to check if a quartic form belongs to Σ2n,4
⋂Sn
4
cvx.
Proof. Let G = (V,E) be a graph with V being the set of n vertices and E being the set of edges.
Define the following bi-quadratic form associated with graph G as follows:
bG(x, y) := 2∑
(i,j)∈E
xixjyiyj .
Ling et al. [24] showed that the problem max‖x‖2=‖y‖2=1 bG(x, y) is strongly NP-hard. Define
bG,λ(x, y) := λ(xTx)(yTy)− bG(x, y) = λ(xTx)(yTy)− 2∑
(i,j)∈E
xixjyiyj .
Then determining the nonnegativity of bG,λ(x, y) in (x, y) is also strongly NP-hard, due to the fact
that the problem max‖x‖2=‖y‖2=1 bG(x, y) can be polynomially reduced to it. Let us now construct
a quartic form in (x, y) as
fG,λ(x, y) := bG,λ(x, y) + n2
n∑i=1
xi4 +
n∑i=1
yi4 +
∑1≤i<j≤n
xi2xj
2 +∑
1≤i<j≤nyi
2yj2
.
Observe that
fG,λ(x, y) = gG,λ(x, y) +∑
(i,j)∈E
(xixj − yiyj)2 + (n2 − 1)∑
(i,j)∈E
(xi2xj
2 + yi2yj
2),
where gG,λ(x, y) := λ(xTx)(yTy) + n2(∑n
i=1(xi4 + yi
4) +∑
(i,j)/∈E(xi2xj
2 + yi2yj
2))
. Therefore
fG,λ(x, y) is quartic SOS in (x, y). Moreover, according to Theorem 2.3 of [2] with γ = 2, we know
that fG,λ(x, y) is convex if and only if bG,λ(x, y) is nonnegative. The latter being strongly NP-hard,
therefore checking the convexity of the quartic SOS form fG,λ(x, y) is also strongly NP-hard. �
21
Figure 1: Hierarchy for the cones of quartic forms
With the help of Theorem 5.4 and Proposition 5.1, which claim that Σ2n,4
⋂Sn
4
cvx is a hard cone while
Σ2∇2
n,4is easy, we conclude the following complete hierarchical structure to clarify one last containing
relationship in Theorem 4.1. (Note that Theorem 4.1 only concludes Σ2∇2
n,4⊆(Σ2n,4
⋂Sn
4
cvx
).)
Corollary 5.5 Assuming P 6= NP , for general n we have
Σ4n,4 ( Sn
2×n2
+ ( Σ2∇2
n,4((Σ2n,4
⋂Sn
4
cvx
)( Σ2
n,4 ( Sn4
+ . (12)
The relationship among these six cones of quartic forms is depicted in Figure 1, where a primal-dual
pair is painted by the same color. The chain of containing relationship is useful especially when
some of the cones are hard while others are easy. One obvious possible application is to use an
‘easy’ cone either as restriction or as relaxation of a hard one. Such scheme is likely to be useful in
the design of approximation algorithms.
5.2 The Low Dimensional Cases
The chain of containing relations (12) holds for general dimension n. Essentially the strict contain-
ing relations are true for n ≥ 4, except that we do not know if Σ2∇2
n,4((Σ2n,4
⋂Sn
4
cvx
)holds true or
not. To complete the picture, in this subsection we discuss quartic forms in low dimensional cases:
n = 2 and n = 3. Specifically, when n = 2, the six cones of quartic forms reduce to two distinctive
ones; while n = 3, they reduce to three distinctive cones.
22
Proposition 5.6 For the cone of bi-variate quartic forms, it holds that
Σ42,4 = S22×22
+ = Σ2∇2
2,4=(Σ2
2,4
⋂S24
cvx
)( Σ2
2,4 = S24
+ .
Proof. By a well known equivalence between nonnegative polynomial and sum of squares due to
Hilbert [19] (for bi-variate quartic polynomials), we have Σ2n,4 = Sn
4
+ for n ≤ 3, by noticing that
Hilbert’s result is true for inhomogeneous polynomials and our cones are for homogeneous forms.
Now, the duality relationship in Theorem 3.5 leads to Σ42,4 = S22×22
+ . Next let us focus on the
relationship between S22×22
+ and Σ22,4
⋂S24cvx. In fact we shall prove below that S24
cvx ⊆ S22×22
+ , i.e.,
any bi-variate convex quartic form is quartic matrix PSD.
For bi-variate convex quartic form F with
F1111 = a1,F1112 = a2,F1122 = a3,F1222 = a4,F2222 = a5,
we have f(x) = F(x, x, x, x) = a1x14 + 4a2x1
3x2 + 6a3x12x2
2 + 4a4x1x23 + a5x2
4, and
∇2f(x) = 12
[a1x1
2 + 2a2x1x2 + a3x22 a2x1
2 + 2a3x1x2 + a4x22
a2x12 + 2a3x1x2 + a4x2
2 a3x12 + 2a4x1x2 + a5x2
2
]� 0 ∀x1, x2 ∈ R. (13)
Denote A1 =
[a1 a2
a2 a3
], A2 =
[a2 a3
a3 a4
]and A3 =
[a3 a4
a4 a5
], and (13) is equivalent to
[xTA1x xTA2x
xTA2x xTA3x
]� 0 ∀x ∈ R2. (14)
According to Theorem 4.8 and the subsequent discussions in [26], it follows that (14) is equivalent
to
[A1 A2
A2 A3
]� 0. Therefore,
F(X,X) = (vec (X))T
[A1 A2
A2 A3
]vec (X) ≥ 0 ∀X ∈ R22 ,
implying that F is quartic matrix PSD. This proves S22×22
+ = Σ22,4
⋂S24cvx. Finally, Example 4.1
for Σ22,4 \ S24
cvx leads to Σ22,4
⋂S24cvx 6= Σ2
2,4. �
It remains to consider the case n = 3. Our previous discussion concluded that Σ23,4 = S34
+ , and so
by duality Σ43,4 = S32×32
+ . Moreover, in a recent working paper Ahmadi et al. [1] showed that every
tri-variate convex quartic polynomial is sos-convex, implying Σ2∇2
3,4=(Σ2
3,4
⋂S34cvx
). So we have
at most three distinctive cones of quartic forms. Example 4.1 in Σ22,4 \ S24
cvx and Example 4.2 in
Σ2∇2
3,4\ S32×32
+ show that there are in fact three distinctive cones.
23
Proposition 5.7 For the cone of tri-variate quartic forms, it holds that
Σ43,4 = S32×32
+ ( Σ2∇2
3,4=(Σ2
3,4
⋂S34
cvx
)( Σ2
3,4 = S34
+ .
6 Quartic Conic Programming
The study of quartic forms in the previous sections gives rise some new modeling opportunities. In
this section we shall discuss quartic conic programming, i.e., optimizing a linear function over the
intersection of an affine subspace and a cone of quartic forms. In particular, we shall investigate
the following quartic conic programming model:
(QCP ) max C • Xs.t. Ai • X = bi, i = 1, . . . ,m
X ∈ Σ4n,4,
where C,Ai ∈ Sn4
and bi ∈ R for i = 1, . . . ,m. As we will see later, a large class of non-convex
quartic polynomial optimization models can be formulated as a special class of (QCP ). In fact
we will study a few concrete examples to show the modeling power of the quartic forms that we
introduced.
6.1 Quartic Polynomial Optimization
Quartic polynomial optimization received much attention in the recent years; see e.g., [25, 24, 15,
16, 35, 23]. Essentially, all the models studied involve optimization of a quartic polynomial function
subject to some linear and/or homogenous quadratic constraints, including spherical constraints,
binary constraints, the intersection of co-centered ellipsoids, and so on. Below we consider a very
general quartic polynomial optimization model:
(P ) max p(x)
s.t. (ai)Tx = bi, i = 1, . . . ,m
xTAjx = cj , j = 1, . . . , l
x ∈ Rn,
where p(x) is a general inhomogeneous quartic polynomial function.
We first homogenize p(x) by introducing a new homogenizing variable, say xn+1, which is set to
one, and get a homogeneous quartic form
p(x) = F(x, x, x, x) = F • (x⊗ x⊗ x⊗ x) ,
24
where F ∈ S(n+1)4 , x =
(x
xn+1
)and xn+1 = 1. By adding some redundant constraints, we have
an equivalent formulation of (P ):
max F (x, x, x, x)
s.t. (ai)Tx = bi,((ai)Tx
)2= bi
2,((ai)Tx
)4= bi
4, i = 1, . . . ,m
xTAjx = cj ,(xTAjx
)2= cj
2, j = 1, . . . , l
x =
(x
1
)∈ Rn+1.
The objective function of the above problem can be taken as a linear function of x⊗ x⊗ x⊗ x, and
we introduce new variables of a super-symmetric fourth order tensor X ∈ S(n+1)4 . The notations
x, X, and X extract part of the entries of X , which are defined as:
x ∈ Rn xi = Xi,n+1,n+1,n+1 ∀ 1 ≤ i ≤ n,X ∈ Sn
2Xi,j = Xi,j,n+1,n+1 ∀ 1 ≤ i, j ≤ n,
X ∈ Sn4 Xi,j,k,` = Xi,j,k,` ∀ 1 ≤ i, j, k, ` ≤ n.
Essentially they can be treated as linear constraints on X . Now by taking X = x ⊗ x ⊗ x ⊗ x,
X = x⊗ x⊗ x⊗ x, and X = x⊗ x, we may equivalently represent the above problem as a quartic
conic programming model with a rank-one constraint:
(Q) max F • Xs.t. (ai)Tx = bi, (ai ⊗ ai) •X = bi
2, (ai ⊗ ai ⊗ ai ⊗ ai) • X = bi4, i = 1, . . . ,m
Aj •X = cj , (Aj ⊗Aj) • X = cj2, j = 1, . . . , l
Xn+1,n+1,n+1,n+1 = 1, X ∈ Σ4n+1,4, rank(X ) = 1.
Dropping the rank-one constraint, we obtain a relaxation problem, which is exactly in the form of
quartic conic program (QCP ):
(RQ) max F • Xs.t. (ai)Tx = bi, (ai ⊗ ai) •X = bi
2, (ai ⊗ ai ⊗ ai ⊗ ai) • X = bi4, i = 1, . . . ,m
Aj •X = cj , (Aj ⊗Aj) • X = cj2, j = 1, . . . , l
Xn+1,n+1,n+1,n+1 = 1, X ∈ Σ4n+1,4.
Interestingly, the relaxation from (Q) to (RQ) is not lossy; or, to put it differently, (RQ) is a tight
relaxation of (Q), under some mild conditions.
Theorem 6.1 If Aj ∈ Sn2
+ for all 1 ≤ j ≤ l in the model (P ), then (RQ) is equivalent to (P ) in
the sense that: (i) they have the same optimal value; (ii) if X is optimal to (RQ), then x is in the
convex hull of the optimal solution of (P ). Moreover, the minimization counterpart of (P ) is also
equivalent to the minimization counterpart of (RQ).
25
This result shows that (P ) is in fact a conic quartic program (QCP ) when the matrices Aj ’s in
(P ) are positive semidefinite. Notice that the model (P ) actually includes quadratic inequality
constraints xTAjx ≤ cj as its subclasses, for one can always add a slack variable yj ∈ R with
xTAjx+ yj2 = cj , while reserving the new data matrix
[Aj 0
0 1
]in the quadratic term still being
positive semidefinite. The proof of Theorem 6.1 is dedicated to Appendix A.
As mentioned before, Burer [8] established the equivalence between a large class of mixed-binary
quadratic programs and co-positive programs. Theorem 6.1 may be regarded as a quartic extension
of Burer’s result. The virtue of this equivalence is to alleviate the highly non-convex objective
and/or constraints of (QCP ) and retain the problem in convex form, although the difficulty is all
absorbed into the dealing of the quartic cone, which is nonetheless a convex one. Note that this is
characteristically a property for polynomial of degree higher than 2: the SDP relaxation for similar
quadratic models can never be tight.
6.2 Biquadratic Assignment Problems
The biquadratic assignment problem (BQAP ) is a generalization of the quadratic assignment prob-
lem (QAP ), which is to minimize a quartic polynomial of an assignment matrix:
(BQAP ) min∑
1≤i,j,k,`,s,t,u,v≤nAijk`BstuvXisXjtXkuX`v
s.t.∑n
i=1Xij = 1, j = 1, . . . , n∑nj=1Xij = 1, i = 1, . . . , n
Xij ∈ {0, 1}, i, j = 1, . . . , n
X ∈ Rn2,
where A,B ∈ Rn4. This problem was first considered by Burkard et al. [10] and was shown to
have applications in the VLSI synthesis problem. After that, several heuristics for (BQAP ) were
developed by Burkard and Cela [11], and Mavridou et al. [27].
In this subsection we shall show that (BQAP ) can be formulated as a quartic conic program
(QCP ). First notice that the objective function of (BQAP ) is a fourth order polynomial function
with respect to the variables Xij ’s, where X is taken as an n2-dimensional vector. The assignment
constraints∑n
i=1Xij = 1 and∑n
j=1Xij = 1 are clearly linear equality constraints. Finally by
imposing a new variable x0 ∈ R, and the binary constraints Xij ∈ {0, 1} is equivalent to(Xij
x0
)T[
1 −1
−1 1
](Xij
x0
)=
1
4and x0 =
1
2,
where the coefficient matrix in the quadratic term is indeed positive semidefinite. Applying Theo-
rem 6.1 we have the following result:
26
Corollary 6.2 The biquadratic assignment problem (BQAP ) can be formulated as a quartic conic
program (QCP ).
6.3 Eigenvalues of Fourth Order Super-Symmetric Tensor
The notion of eigenvalue for a matrix has been extended to tensors; see e.g., [30, 31, 32]. Versatile
extensions turned out to be possible, among which Qi [30] proposed the following notion of the
Z-eigenvalues: Restricting to the space of fourth order super-symmetric tensors Sn4, λ ∈ R is called
Z-eigenvalue of the super-symmetric tensor F ∈ Sn4, if the following system holds{
F(x, x, x, ·) = λx,
xTx = 1,
where x ∈ Rn is the corresponding eigenvector with respect to λ. Notice that the Z-eigenvalues are
the usual eigenvalues for a symmetric matrix, when restricting to the space of symmetric matrices
Sn2. We refer interested readers to [31] for various properties of Z-eigenvalue and [32] for its
applications in polynomial optimizations.
Observe that x is a Z-eigenvector of the fourth order tenor F if and only if x is a KKT point to
following polynomial optimization problem:
(E) max F(x, x, x, x)
s.t. xTx = 1.
Furthermore, x is the Z-eigenvector with respect to the largest (resp. smallest) Z-eigenvalue of Fif and only if x is optimal to (E) (resp. the minimization counterpart of (E)). As the quadratic
constraint xTx = 1 satisfies the condition in Theorem 6.1, we reach the following conclusion:
Corollary 6.3 The problem of finding a Z-eigenvector with respect to the largest or smallest Z-
eigenvalue of a fourth order super-symmetric tensor F can be formulated as a quartic conic program
(QCP ).
To conclude this section, as well as the whole paper, we remark here that quartic conic problems have
many potential application, alongside their many intriguing theoretical properties. The hierarchical
structure of the quartic cones that we proved in the previous sections paves a way for possible
relaxation methods to be viable. For instance, according to the hierarchy relationship (12), by
relaxing the cone Σ4n,4 to an easy cone Sn
2×n2
+ lends a hand to solve the quartic conic problem
approximately. The quality of such solution methods and possible enhancements remain our future
research topic.
27
References
[1] A.A. Ahmadi, G. Blekherman, and P.A. Parrilo, Every Convex Ternary Quartic is SOS-
Convex, Working Paper, Department of Electrical Engineering and Computer Science, Mas-
sachusetts Institute of Technology, 2011. 18, 23
[2] A.A. Ahmadi, A. Olshevsky, P.A. Parrilo, and J.N. Tsitsiklis, NP-hardness of Deciding
Convexity of Quartic Polynomials and Related Problems, Preprint, December 2010. http:
//arxiv.org/abs/1012.1908v1 1, 2, 3, 21
[3] A.A. Ahmadi and P.A. Parrilo, A Convex Polynomial that Is Not SOS-Convex, Mathematical
Programming, Series A, to appear. 3, 7, 16, 17
[4] A.A. Ahmadi and P.A. Parrilo, On the Equivalence of Algebraic Conditions for Convexity and
Quasiconvexity of Polynomials, Proceedings of the 49th IEEE Conference on Decision and
Control, 2010. 7
[5] A. Barvinok, A Course in Convexity, Graduate Studies in Mathematics Volume 54, American
Mathematical Society, 2002. 12
[6] G. Blekherman, Convex Forms That Are Not Sums of Squares, Preprint, October 2009. http:
//arxiv.org/abs/0910.0656 3, 18
[7] P. Biswas, T.-C. Liang, T.-C. Wang, and Y. Ye, Semidefinite Programming Based Algorithms
for Sensor Network Localization, ACM Transactions on Sensor Networks, 2, 188–220, 2006. 2,
4
[8] S. Burer, On the Copositive Representation of Binary and Continuous Nonconvex Quadratic
Programs, Mathematical Programming, Series A, 120, 479–495, 2009. 20, 26, 33
[9] S. Burer and H. Dong, Representing Quadratically Constrained Quadratic Programs as Gener-
alized Copositive Programs, Technical Report. http://www.optimization-online.org/DB_
FILE/2011/07/3095.pdf 20
[10] R.E. Burkard, E. Cela and B. Klinz, On the biquadratic assignment problem. In: Pardalos,
P.M., Wolkowicz, H. (Eds.), Quadratic Assignment and Related Problems, DIMACS Series
in Discrete Mathematics and Theoretical Computer Science, Vol. 16. AMS, Rhode Island,
117–146, 1994. 26
[11] R.E. Burkard and E. Cela, Heuristis for Biqudratic Assignment Problems and Their Compu-
tational Comparison, European Journal of Operational Research 83, 283–300, 1995. 26
28
[12] C. Chen and P.P. Vaidyanathan, MIMO Radar Waveform Optimization With Prior Informa-
tion of the Extended Target and Clutter, IEEE Transactions on Signal Processing, 57, 3533–
3544, 2009. 2
[13] G. Dahl, J.M. Leinaas, J. Myrheim, and E. Ovrum, A Tensor Product Matrix Approximation
Problem in Quantum Physics, Linear Algebra and Applications, 420, 711–725, 2007. 2
[14] P.J.C. Dickinson and L. Gijben, On the Computational Complexity of Membership Prob-
lems for the Completely Positive Cone and Its Dual, Technical Report. http://www.
optimization-online.org/DB_HTML/2011/05/3041.html 19
[15] S. He, Z. Li, and S. Zhang, Approximation Algorithms for Homogeneous Polynomial Optimiza-
tion with Quadratic Constraints, Mathematical Programming, Series B, 125, 353–383, 2010.
2, 24
[16] S. He, Z. Li, and S. Zhang, General Constrained Polynomial Optimization: An Approximation
Approach, Technical Report SEEM2009-06, Department of Systems Engineering & Engineering
Management, The Chinese University of Hong Kong, 2009. 2, 24
[17] S. He, B. Jiang, Z. Li, and S. Zhang, A Note on Hilbert’s Identity, Technical Report SEEM2011-
01, Department of Systems Engineering & Engineering Management, The Chinese University
of Hong Kong, 2011. 12
[18] J.W. Helton and J. Nie, Semidefinite Representation of Convex Sets, Mathematical Program-
ming, Series A, 122, 21–64, 2010. 3, 7, 8
[19] D. Hilbert, Uber die Darstellung Definiter Formen als Summe von Formenquadraten, Mathe-
matische Annalen, 32, 342–350, 1888. 3, 23
[20] P.M. Kleniati, P. Parpas, and B. Rustem, Partitioning Procedure for Polynomial Optimization:
Application to Portfolio Decisions with Higher Order Moments, COMISEF Working Papers
Series, WPS-023, 2009. 2, 3
[21] J.B. Lasserre, Global Optimization with Polynomials and the Problem of Moments, SIAM
Journal on Optimization, 11, 769–817, 2001. 2
[22] M. Laurent, Sums of Squares, Moment Matrices and Optimization over Polynomials, in: Puti-
nar, M., Sullivant, S. (eds.) Emerging Applications of Algebraic Geometry, Series: The IMA
Volumes in Mathematics and its Applications, vol. 149, Springer, Berlin, 2009. 2
[23] Z. Li, Polynomial Optimization Problems—Approximation Algorithms and Applications, Ph.D.
Thesis, The Chinese Univesrity of Hong Kong, June 2011. 2, 24
29
[24] C. Ling, J. Nie, L. Qi, and Y. Ye, Biquadratic Optimization Over Unit Spheres and Semidefinite
Programming Relaxations, SIAM Journal on Optimization, 20, 1286–1310, 2009. 2, 3, 21, 24
[25] Z.-Q. Luo and S. Zhang, A Semidefinite Relaxation Scheme for Multivariate Quartic Polynomi-
al Optimization With Quadratic Constraints, SIAM Journal on Optimization, 20, 1716–1736,
2010. 2, 24
[26] Z.-Q. Luo, J.F. Sturm, and S. Zhang, Multivariate Nonnegative Quadratic Mappings, SIAM
Journal on Optimization, 14, 1140–1162, 2004. 2, 23
[27] T. Mavridou, P.M. Pardalos, L.S. Pitsoulis and M.G.C. Resende, A GRASP for the biquadratic
assignment problem, European Journal of Operational Research 105, 613–621, 1998. 26
[28] P.M. Pardalos and S.A. Vavasis, Open Questions in Complexity Theory for Numerical Opti-
mization, Mathematical Programming, 57, 337–339, 1992. 2, 3
[29] P.A. Parrilo, Structured Semidefinite Programs and Semialgebraic Geometry Methods in Ro-
bustness and Optimization, PhD Dissertation, California Institute of Technology, 2000. 2
[30] L. Qi, Extrema of a Real Polynomial, Journal of Global Optimization, 30, 405–433, 2004. 27
[31] L. Qi, Eigenvalues of a Real Supersymmetric Tensor, Journal of Symbolic Computation, 40,
1302–1324, 2005. 27
[32] L. Qi, F. Wang and Y. Wang, Z-eigenvalue Methods for a Global Polynomial Optimization
Problem, Mathematical Programming, Series A, 118, 301–316, 2009. 27
[33] B. Reznick, Some Concrete Aspects of Hilbert’s 17th Problem, Real Algebraic Geometry and
Ordered Structures, Contemporary Mathematics, 253, American Mathematical Society, 2000.
15
[34] B. Reznick, Some of Even Powers of Real Linear Forms, Memoirs of the American Mathe-
matical Society, Volume 96, Number 463, March 1992. 3, 13
[35] A. M.-C. So, Deterministic Approximation Algorithms for Sphere Constrained Homogeneous
Polynomial Optimization Problems, Mathematical Programming, Series B, to appear. 2, 24
[36] J.F. Sturm and S. Zhang, On Cones of Nonnegative Quadratic Functions, Mathematics of
Operations Research, 28, 246–267, 2003. 2, 3, 9, 13, 18
30
A Proof of Theorem 6.1
Here we only prove the equivalence relation for the maximization problems, since the proof for their
minimization counterparts is exactly the same. That is, we shall prove the equivalence between
(Q) and (RQ).
To start with, let us first investigate the feasible regions of these two problems, to be denoted by
feas (Q) and feas (RQ) respectively. The relationship between feas (Q) and feas (RQ) is revealed by
the following lemma.
Lemma A.1 It holds that conv (feas (Q)) ⊆ feas (RQ) = conv (feas (Q)) + P, where
P := cone
{(x
0
)⊗(x
0
)⊗(x
0
)⊗(x
0
)∣∣∣∣ (ai)Tx = 0 ∀ 1 ≤ i ≤ m,xTAjx = 0 ∀ 1 ≤ j ≤ l
}⊂ Σ4
n+1,4.
Proof. First, it is obvious that conv (feas (Q)) ⊆ feas (RQ), since (RQ) is a relaxation of (Q) and
feas (RQ) is convex. Next we notice that the recession cone of feas (RQ) is equal toX ∈ Σ4n+1,4
∣∣∣∣∣∣∣Xn+1,n+1,n+1,n+1 = 0,
(ai)Tx = 0, (ai ⊗ ai) •X = 0, (ai ⊗ ai ⊗ ai ⊗ ai) • X = 0 ∀ 1 ≤ i ≤ m,Aj •X = 0, (Aj ⊗Aj) • X = 0 ∀ 1 ≤ j ≤ l
.
Observing that X ∈ Σ4n+1,4 and Xn+1,n+1,n+1,n+1 = 0, it is easy to see that x = 0 and X = 0. Thus
the recession cone of feas (RQ) is reduced toX ∈ Σ4n+1,4
∣∣∣∣∣∣∣Xn+1,n+1,n+1,n+1 = 0, x = 0, X = 0,
(ai ⊗ ai ⊗ ai ⊗ ai) • X = 0 ∀ 1 ≤ i ≤ m,(Aj ⊗Aj) • X = 0 ∀ 1 ≤ j ≤ l
⊇ P,
which proves feas (RQ) ⊇ conv (feas (Q)) + P.
Finally, we shall show the inverse inclusion, i.e., feas (RQ) ⊆ conv (feas (Q)) + P. Suppose X ∈feas (RQ), then it can be decomposed as
X =∑k∈K
(yk
αk
)⊗(yk
αk
)⊗(yk
αk
)⊗(yk
αk
), (15)
where αk ∈ R, yk ∈ Rn for all k ∈ K. Immediately we have∑k∈K
αk4 = Xn+1,n+1,n+1,n+1 = 1. (16)
31
Now divide the index set K into two parts, with K0 := {k ∈ K | αk = 0} and K1 := {k ∈ K | αk 6=0}, and let zk = yk/αk for all k ∈ K1. The decomposition (15) is then equivalent to
X =∑k∈K1
αk4
(zk
1
)⊗(zk
1
)⊗(zk
1
)⊗(zk
1
)+∑k∈K0
(yk
0
)⊗(yk
0
)⊗(yk
0
)⊗(yk
0
).
If we can prove that (zk
1
)⊗(zk
1
)⊗(zk
1
)⊗(zk
1
)∈ feas (Q) ∀ k ∈ K1 (17)(
yk
0
)⊗(yk
0
)⊗(yk
0
)⊗(yk
0
)∈ P ∀ k ∈ K0 (18)
then by (16), we shall have X ∈ conv (feas (Q)) + P, proving the inverse inclusion.
In the following we shall prove (17) and (18). Since X ∈ feas (RQ), together with x =∑
k∈K αk3yk,
X =∑
k∈K αk2yk ⊗ yk, and X =
∑k∈K y
k ⊗ yk ⊗ yk ⊗ yk, we obtain the following equalities:∑k∈K
αk3(ai)Tyk = bi,
∑k∈K
αk2(
(ai)Tyk)2
= bi2,∑k∈K
((ai)Tyk
)4= bi
4, ∀ 1 ≤ i ≤ m
∑k∈K
αk2(yk)TAjyk = cj ,
∑k∈K
((yk)TAjyk
)2= cj
2, ∀ 1 ≤ j ≤ l.
As a direct consequence of the above equalities and (16), we have(∑k∈K
αk2 · αk(ai)Tyk
)2
= bi2 =
(∑k∈K
αk4
)(∑k∈K
αk2(
(ai)Tyk)2), ∀ 1 ≤ i ≤ m
(∑k∈K
αk2(
(ai)Tyk)2)2
= bi4 =
(∑k∈K
αk4
)(∑k∈K
((ai)Tyk
)4), ∀ 1 ≤ i ≤ m
(∑k∈K
αk2(yk)TAjyk
)2
= cj2 =
(∑k∈K
αk4
)(∑k∈K
((yk)TAjyk
)2), ∀ 1 ≤ j ≤ l.
Noticing that the equalities hold for the above Cauchy-Schwarz inequalities, it follows that for every
1 ≤ i ≤ m and every 1 ≤ j ≤ l, there exist δi, εi, θj ∈ R, such that
δiαk2 = αk(a
i)Tyk, εiαk2 =
((ai)Tyk
)2and θjαk
2 = (yk)TAjyk ∀ k ∈ K. (19)
If αk = 0, then (ai)Tyk = 0 and (yk)TAjyk = 0, which implies (18). Moreover, due to (19) and (16),
δi = δi
(∑k∈K
αk4
)=∑k∈K
δiαk2 · αk2 =
∑k∈K
αk(ai)Tyk · αk2 = bi ∀ 1 ≤ i ≤ m.
32
Similarly, we have θj = cj for all 1 ≤ j ≤ l. If αk 6= 0, noticing zk = yk/αk, it follows from (19)
that(ai)Tzk = (ai)Tyk/αk = δi = bi ∀ 1 ≤ i ≤ m
(zk)TAjzk = (yk)TAjyk/αk2 = θj = cj ∀ 1 ≤ j ≤ l,
which implies (17). �
To prove the original Theorem 6.1, we notice that if Aj is positive semidefinite, then xTAjx =
0⇐⇒ Ajx = 0. Therefore,
(x
0
)⊗(x
0
)⊗(x
0
)⊗(x
0
)∈ P implies that x is a recession direction of
the feasible region for (P ). With this property and using a similar argument of Theorem 2.6 in [8],
Theorem 6.1 follows immediately.
33